Anders Hejlsberg on the worst programming language for AI

Key Concepts

• AI Code Generation: The ability of Artificial Intelligence to produce computer code.
• Training Data: The dataset used to teach an AI model, in this case, existing code in various programming languages.
• Extrapolation: The process by which an AI model generates new code based on patterns learned from the training data.
• Regurgitation (in the context of AI): The tendency of AI models to reproduce patterns and structures directly observed in the training data.
• Data Dependency: The performance of AI code generation is heavily reliant on the quantity and quality of training data.

The Paradox of the "Perfect" AI Programming Language

The central question addressed is whether a specifically designed programming language would be optimal for AI code generation. The speaker argues against this notion, presenting a counterintuitive perspective: the most effective language for AI is not a novel, perfectly designed one, but rather a language the AI has been extensively exposed to during its training.

The core argument rests on the fundamental mechanism of how current AI models generate code. They function as “big regurgitators,” meaning their ability to write code in a specific language is directly proportional to the amount of code in that language present in their training dataset. This isn’t about understanding the underlying logic of programming, but about recognizing and reproducing patterns. The speaker clarifies this with the statement, “AI’s abilities in a particular writing code in a particular language is directly proportional to how much of that code it has seen because it is a big regurgitator if you will of stuff that someone has done with some extrapolation on on top of it.”

Evidence from Existing Data & Language Popularity

The speaker supports this claim with empirical observation. Data indicates that AI models demonstrate proficiency in languages like JavaScript, Python, and TypeScript. This proficiency isn’t due to inherent qualities of these languages, but because these languages are heavily represented in the vast amounts of code the AI has been trained on. The phrase “we we know from the data right that AI has seen a lot of JavaScript and a lot of Python and a lot of TypeScript and therefore it's pretty darn good at writing code in those languages” highlights this direct correlation.

Implications for New Programming Languages

This has significant implications for the development of new programming languages intended for AI use. The speaker concludes that new languages are, in fact, disadvantaged in this context. Because they lack a substantial body of existing code for AI models to learn from, they will initially perform poorly in code generation tasks.

Logical Flow & Synthesis

The argument progresses logically from posing the initial question about a “perfect” AI language, to explaining the underlying mechanism of AI code generation (pattern recognition and reproduction), to providing supporting evidence based on the performance of AI in popular languages, and finally, to drawing a conclusion about the challenges faced by new languages.

The main takeaway is that, currently, the best language for AI to target isn’t about linguistic elegance or theoretical optimality, but about sheer volume of available training data. This suggests that focusing on widespread adoption and code generation in existing languages is a more effective strategy than attempting to create a new language specifically for AI.